Sex and Age Estimation from Cardiac Signals Captured via Radar using Data Augmentation and Deep Learning: A Privacy Concern

Daniel Foronda-Pascual^*Carmen CamaraPedro Peris-Lopez

• Universidad Carlos III de Madrid de Madrid, Leganés, Spain

Introduction Electrocardiograms (ECGs) have long served as the standard method for cardiac monitoring. While ECGs are highly accurate and widely validated, they require direct skin contact, are sensitive to motion artifacts, and are not always practical for continuous or unobtrusive monitoring, limiting their generalization to real-world, dynamic environments. However, radar-based technologies offer a novel, non-invasive alternative for acquiring cardiac signals without direct contact. This improves both hygiene and patient comfort, making it especially attractive for medical applications. Despite these benefits, it may raise privacy concerns, inadvertently revealing personal attributes such as sex and age. This study investigates, for the first time, how such demographic information can be inferred from radar-acquired cardiac signals. Methods To address this question, we developed a machine learning framework to predict demographic attributes from radar-based cardiac signals. These signals were transformed into scalograms—a time-frequency representation—and then classified using a Convolutional Neural Network (CNN). Given the lack of prior studies on demographic inference from radar-based cardiac signals, the generalization capabilities of existing approaches remain untested in this context. Moreover, the small size of available datasets further limits model performance. To mitigate these issues, we applied data augmentation using a Conditional Wasserstein Generative Adversarial Network (cWGAN), which generated synthetic scalograms conditioned on class labels. Notably, there are very few prior studies on data augmentation specifically for this type of signal. This strategy aimed to enhance model accuracy and generalization by enriching the training data. Results Our experiments demonstrate that data augmentation significantly improves model performance. The trained model achieved an accuracy of 78.40% in predicting the sex of individuals and 72.83% accuracy in classifying them into two age groups (18–29 and 30–65 years), despite the dataset being limited to only 30 subjects. Discussion These findings reveal a potential privacy risk associated with radar-based biometric systems. The ability to infer sensitive demographic information from physiological signals could have serious implications, particularly in secure applications such as electronic passports (e-passports), where access to RFID chip data often depends on such personal attributes. Therefore, while radar technologies offer promising advantages, their deployment must consider and address the associated privacy challenges.